The release of ChatGPT to the general public in November 2022 marked the beginning of a new era in the way we interact with machines. Suddenly, it seems, computers are able to understand natural language. No longer mechanically and rule based as before, but intuitively, nuanced and eloquent like humans. Even better.

This revolution is powered by large language models (LLMs). These models use artificial neural networks with billions of parameters, allowing them to understand and generate human-like text. To put this into perspective, traditional algorithms like those used in basic data analysis have only a few dozen to a few hundred parameters.

Since LLMs became accessible, a lively debate has emerged among IP professionals, mirroring discussions in other sectors. The pressing questions are: Should AI tools be used? To what extent? Will they replace us in the near future?

The Role of AI in Patent Law

Patent attorneys primarily work with texts and language. LLMs do too. It’s hard to believe that LLMs won’t disrupt our profession.

Of course, this debate also concerns other industries. In programming, the development is compared with that of the autonomous vehicle, which can also be applied to the patent industry:

Level 1: AI as an Assistive System

At this level, AI helps patent attorneys to work more efficiently and accurately. It serves as a tool that simplifies patent searches, analyzes patent documents in detail, and automates parts of the drafting process. This not only improves the quality of patent applications but also reduces the time required for document creation and review.

Level 2: AI as an Autonomous Agent

AI could take on more active roles, such as carrying out freedom-to-operate searches, independently drafting patent claims or compiling entire patent applications based on technical descriptions provided by developers or inventors. In this scenario, AI collaborates with humans, and performs specific tasks under supervision of patent attorneys. AI systems could also take on the role of a quality manager, for example by performing internal consistency checks to ensure that all required information is accurately presented and protected.

Level 3: Complete Takeover by AI

In the third and most radical stage, AI could almost entirely take over the role of the patent attorney. It would not only perform the drafting but also make strategic decisions, such as the design of scope and protection according to strategic guidelines, and the management of objections and legal disputes. At this level, patent attorneys might focus more on advisory functions or on monitoring of the work created by AI.

Current and Future Impact of AI

While the final stage of AI autonomy is intriguing and holds significant societal implications, it remains a distant prospect. However, we have reached at least level 1. And this has already the potential to fundamentally change our work.

To remain relevant, it is essential for human intellectual property workers to understand the technology behind AI. What are the possibilities? What are the limitations?

I recently wrote a LinkedIn post critically examining the use of ChatGPT by patent attorneys. Though I am generally a strong advocate for using language models in our profession, it is not wise to use the most popular ones like ChatGPT for every task. Depending on the case, there may be legal obstacles, confidentiality issues or there are better solutions to a specific problem, a specialized AI/LLM or even non-AI.

A brief example from practice may illustrate this point. When drafting patent applications, besides the – analytical and even creative – work, drafting involves a huge amount of rule-based routine work as well. Such as adding reference signs or mirroring the claim language to the description. This is hardly an outstanding use case for (cloud-based) LLMs. There are plenty of existing conventional tools that can already perform these functions more reliably – and locally.

Alternatively, why not quickly create your own tool – and why not even with ChatGPT’s help? For example, “Write a short Python script to add reference signs to a patent description”.

Legal and Functional Limitations of LLMs in IP Law

Legal Restrictions

As professional secrecy holders, at least in Germany, patent attorneys face strict restrictions regarding the use of cloud software. Currently, it seems often not possible to meet all legal requirements for using many cloud services with confidential data. Even if there were no specific legal hurdles, the extent to which client’s confidential information is sent to third-party companies should be carefully considered.

One solution is local open-source LLMs, which are constantly improving and can be better customized to specific needs through fine-tuning.

Functional Limitations

The primary function of LLMs today is language comprehension based on predicting answers to questions. While useful, this is not comparable to a thinking process or the full spectrum of human intelligence.

For example, drafting a patent application involves more than technical understanding and transforming the invention report into a legal text. It requires considering various stakeholders and aspects, including the patent department, inventors, product managers, corporate strategy, legal framework, case law, and more. This goes far beyond what LLMs like ChatGPT are trained for.

It quickly becomes clear that the weakness of current AI models is their limited ability and bandwidth to holistically collect, process and analyze the necessary input to the required extent. This is where humans come in.

So, it depends on the input. Good patent attorneys can achieve even better results using LLMs if they use it with well-designed prompts or tools. Conversely, users without sufficient legal skills may produce patent applications that read convincingly but are still of poor quality and will not stand up to legal proceedings.

In other sectors like programming, it’s a similar situation with GenAI. When LLMs are used to generate code, it always looks too perfect. A good programmer can still recognize weaknesses and fix them. And often the program is not even executable. Unfortunately, we patent attorneys cannot benefit from such immediate feedback. Patent applications are examined after months and weaknesses may even become apparent years later, as in the event of an infringement suit.

Practical Applications and Automation Opportunities

Despite these limitations, I am firmly convinced that the chances of remaining competitive decrease with every day that the use of AI tools is postponed. Of course, this requires digitalization. In this case, the rule is: first digitize, then automate. And: It’s about time you got going!

Here are a few examples how to use AI tools productively already today:

  • Patent Search: Modern concepts such as semantic search and RAG (Retrieval-Augmented Generation) can improve and simplify the search for relevant property rights. AI-based tools with features like “ask-your-data” are particularly useful.
  • Responses to office actions: AI can assist in evaluating cited prior art, searching for case law, or brainstorming better arguments. For published patent applications, ChatGPT should also be usable without any legal troubles.
  • Patent drafting: Interacting with patent documents or invention disclosures via chat using LLMs is fun and can speed up the initial familiarization process. While AI can enrich claim drafting and description supplementation, thorough review of AI-generated texts remains essential.
  • Automatization: AI can also be part of an automatization strategy for many tasks directly in the work of attorneys or in the workflows of a patent law firm.

Summary and Conclusion

The above list is just a small sample of what is already possible with LLM today. Many AI use cases, such as ‘ask-your-data’, not only have the potential, but have already changed the way we work for many.

The key is to begin integrating these tools effectively and to stay competitive in this evolving landscape. One should also not forget a characteristic feature of current AI models: They only achieve good results if the input is excellent. This is where humans come into play again.

About the blogpost author:

Sebastian Goebel is a European and German patent attorney, UPC representative, and co-founder of the patent law firm Bösherz Goebel with a primary focus on innovations in the field of digital technologies.

His professional journey commenced as an electrical engineer and software developer, where he also contributed to research in medical technology and Machine Learning (ML), among others at the Ruhr University Bochum, and the University of California in Los Angeles. In this context, he was also awarded the Inventor Prize of the Ruhr University Bochum. In addition, he holds a Master’s degree in Lasers and Photonics and contributed as a lecturer to the Intellectual Property course at the Ruhr University Bochum.

Based on his practical experience in various technological areas, his expertise lies in the interdisciplinary use of Machine Learning together with other engineering sciences. In addition, he is an AI enthusiast who is passionate about the use of Machine Learning tools in the field of legal work.